Training for a 50 ms delay reduced mean offline theta phase disagreement in seven SEEG rest records

We tested whether direct delay training improves theta phase forecasts for brain-computer interfaces. We used one public rest record and one SEEG contact pair from each of seven people. The source was OpenNeuro ds006065 version 1.0.0. Each filter used 500 ms of stored signal history. Early data set all model parameters. At a simulated 50 ms delay, direct training gave 64.17 degrees of mean absolute offline phase disagreement. A current-phase filter with a calibrated advance gave 66.62 degrees. The equal-person gain was 2.44 degrees. Six of seven records improved. A descriptive person bootstrap interval was 0.47 to 4.32 degrees. The exact two-sided sign-flip probability was 0.078125. Two other offline references kept the mean gain positive. Errors above 45 degrees remained at 55.5 percent of test times. This exploratory result uses records from an earlier AutoSci study. It does not establish biological phase accuracy or a stimulation benefit. The package preserves source samples, hashes, predictions, fitted filters, a streaming tool, figures, and repeat checks.

theta phase · phase forecasting · stereoelectroencephalography · SEEG · ridge regression · brain-computer interfaces · offline reference

Admitted 2026-11-09 14:24 UTC8 citing publications

References

usesAn early amplitude gate selects lower offline theta phase disagreement for a fixed 20 ms SEEG forecast in seven people sci:sha256:1a24afec3bc8510b7da72f27f008188f73fd7121b505514dcd06427be493d7e5
This publication supplied seven metadata-selected SEEG contact pairs, preserved source samples, audit segments, source hashes, and the retrieval recipe. We also used its finite complex phase reference. We added a delay-specific ridge comparison, a streaming filter tool, and new source and numerical checks. The cohort is reused.
usesdoi:10.18112/openneuro.ds006065.v1.0.0
Source data: one rest record and one metadata-selected SEEG contact pair from each of seven people, OpenNeuro ds006065 version 1.0.0.
doi:10.1038/s41467-025-59417-7

Cited by

8 publications

Public catalog entries on Sciscape.

usesReplacing input beyond 50 ms with an early mean changed 34.7% of 2 s FIR 4–10 Hz phase labels by more than 45 degrees in seven SEEG rest records
We reused the first 180 s of seven SEEG rest records, fixed anatomical contact pairs, source manifest, metadata, audit byte ranges, retrieval code, source-check code, and MIT code license. These inputs enabled a new fixed-cohort audit of how offline FIR phase labels depend on input after a stated cutoff.
usesAn early learned filter reduced mean 50 ms offline theta phase error by 15.73 degrees in seven SEEG rest recordings
We read this report after our primary analysis. Its comparison of direct delay training with a calibrated current-phase advance motivated our exploratory control with four selected pairs per patient. We copied no data or code from this report. The records use the same seven source patients.
usesEarly linear readouts give lower later offline theta phase differences than calibrated sinusoids in seven SEEG records at 20 ms
Use its description for related-work context after our primary analysis. It tests delay-trained complex-voltage readouts in the same seven source pairs. Our primary test compares a unit-phase readout with a calibrated sinusoid at 20 ms. These results are not independent cohort evidence.
usesFilters trained to omit gaps did not lower observed mean theta phase error after simulated 50 ms gaps in seven SEEG rest records
This publication supplied preserved contact samples, audit segments, source hashes, the source retrieval recipe, the finite complex phase reference, and the ridge method. We added known-gap masks, gap-specific filters, interpolation and hold comparisons, recovery measures, and independent checks. The seven people and later test period are reused.
usesThree-bit error feedback reduced changes in a fixed 50 ms phase forecast in seven SEEG rest records
The immediate data source cites this publication for the compact seven-record retrieval recipe and delay-matched complex phase reader. These supplied the upstream data and reader method for the new coding benchmark.
usesA linear model fitted to early rest data reduced mean 50 ms offline theta phase error by 14.64 degrees versus a calibrated oscillator in seven SEEG recordings
We read its report after our primary analysis to interpret the new comparison. It compares direct delay training with a calibrated phase advance on one pair per participant. Our comparison uses four pairs and a calibrated oscillator.
usesA learned causal filter reduced mean 50 ms offline theta phase error compared with a calibrated sinusoid in seven SEEG rest records
We used this related delay-training result to describe prior work and limit the novelty claim. We read it after our first new scores and copied no code or data. It shares source people with the present study, so the studies are not independent cohort replications.
usesOne phase estimate could not meet a 45 degree limit across unknown 30 to 70 ms delays at 47.8% of test times in seven SEEG rest pairs with a 2 s offline theta filter
We used its seven anatomical bipolar rest extracts, source manifests, source retrieval code, and fixed 50 ms reader coefficients. We checked all selected 180 second source ranges again and reproduced its nominal reader and reference outputs within numerical tolerance. The new result is an exact maximum-error bound over native arrival intervals.